Click here to


Are you sure ?

Yes, do it No, cancel

Deep Learning in the Task of Detecting Coronary Artery Calcifications On Low-Dose Thoracic CTs

J D Fuhrman1*, R Yip2 , C I Henschke2 , D F Yankelevitz2 , M L Giger1 , (1) Committee on Medical Physics, Department of Radiology, The University of Chicago, Chicago, IL, (2) Department of Radiology, Mount Sinai School of Medicine, New York, NY


(Sunday, 7/14/2019) 2:00 PM - 3:00 PM

Room: Stars at Night Ballroom 2-3

Purpose: To detect coronary artery calcification (CAC), an independent indicator of major cardiac events, in low-dose thoracic CT (LDCT) images using a cascade of convolutional neural networks.

Methods: The dataset included 863 LDCT cases acquired with no contrast or ECG-gating. Images were cropped to a 256 x 256 pixel region surrounding the heart, and 120 randomly selected cases (3,121 images) were annotated for training a cascade of deep networks. The architecture consisted of two 2D U-Nets, one tasked with heart segmentation (the search region) and one tasked with aorta segmentation (for elimination of aortic valve calcifications). Segmentation performance was evaluated using the Dice-Sorensen Coefficient (DSC).Subsequently, calcifications were detected through two methods. For Method 1, the total number of calcified tissue pixels was summed across all slice search regions per case. Alternatively, Method 2 used transfer learning to extract features from the pooling layers of a VGG-19 architecture pre-trained on the ImageNet database, reduced feature dimension through PCA, and classified cases using a SVM with 5-fold cross validation. Pixel counts and SVM output were used as input for ROC analysis, and the area under the ROC curve (AUC) was used as a performance metric in the task of CAC detection.

Results: For all results, the 95% confidence interval is listed. Segmentations of the heart and aorta were performed with average DSCs of 0.953 (0.951,0.955) and 0.793 (0.789,0.797), respectively. Method 1 produced an AUC of 0.755 (0.733,0.777), while Method 2 produced an AUC of 0.788 (0.786,0.790).

Conclusion: The removal of the aorta and tissue outside the heart provided useful results compared to random classification (AUC=0.5, Method 1 & Method 2: p<0.0001) but failed to show a statistically significant difference compared to previous methods (AUC=0.790 (0.760,0.820), p=0.83). Overall, these results show promise for the detection of coronary artery calcifications in LDCT images.

Funding Support, Disclosures, and Conflict of Interest: Supported, in part, by the NIBIB of the NIH under grant number T32 EB002103. MLG is a stockholder in R2 Technology/Hologic, a cofounder and equity holder in Quantitative Insights, and a shareholder in QView. MLG receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi, and Toshiba.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email